Claude Code 2.1.240

• [new] Optional thinking display updates path (beta header thinking-display-updates-2026-08-18), so streaming/thinking UI can surface incremental updates when enabled.
• [changed] Agent guidance refined for focus mode (final message only—no mid-turn progress chatter), corrections → memory (write durable preferences in the same turn you engage the fix), and context management when long threads are summarized.
• [changed] Subagent instructions stress background-by-default behavior, never inventing pending-agent results, and when run_in_background: false is appropriate; remote isolation and teammate limits called out where they apply.
• [changed] Safety and workflow prompts tightened: avoid chat/ticket exfiltration unless directed, careful multi-session git stash handling (tagged stash + apply by SHA, prefer WIP commits), PR follow-up Slack posting only when CLAUDE.md implies it and the user confirms, and browser/computer-use MCP tab handling.
• [new] Configuration knobs for Microsoft Foundry-style setups (ANTHROPIC_FOUNDRY_API_KEY, ANTHROPIC_FOUNDRY_BASE_URL, ANTHROPIC_FOUNDRY_RESOURCE) and related hosting metadata.
• [new] Session and UX controls, including scroll speed, tmux truecolor, mouse disable, session naming, subagent model override, max web searches per session, resume prompt, and OAuth scope configuration.
• [changed] SDK compatibility matrix updated (drops older 0.3.201 test entry; adds 0.3.224).

Models & prompts

• Default Sonnet description can be overridden via environment for custom model labeling.
• Built-in behaviour emphasises autonomous execution with fewer interruptions, fair handling when users reaffirm a request after a concern, and plain correction of real mistakes without re-auditing accurate earlier answers.
• Plan mode, permission-denial recovery, auto-mode classifier review, and task-list update tooling prompts remain in the loop with clearer staging (block-first classifier replies, read-only work still allowed without the classifier).
• Onboarding-guide and high-effort review helper prompts adjusted for power-user / teammate walkthroughs and recall-oriented review.

Under the hood

• New runtime flags for background shell pressure reaping, cron disable, memory periodic resync disable, nonessential traffic disable, loop keepalive, host auth refresh timeout, plugin/skills sync install timeouts, PTY orphan checks, bridge reattach sequencing, WebSocket auth via file descriptor, Perfetto tracing, voice interim forwarding, artifact/container metadata, and git-bash path override.
• Feature gates added for thinking display updates, medium-effort nudges (shown/resolved), memory stream listing, team memory conflict recovery, tool memory cgroup handling, and related internal experiments (tengu_radiant_island, tengu_sable_thrush, and companions); older tab-read and tool-cgroup gates removed.
• API surface gains the thinking-display-updates-2026-08-18 beta; workflow-name-only env removed in favor of newer session naming controls.

Upgrade notes

• No forced behaviour change expected for default installs; thinking display updates and Foundry/session env vars are opt-in.
• If you pin SDK wrapper versions, note the tested set now includes 0.3.224 and no longer lists 0.3.201.
• Custom automation that depended on CLAUDE_WORKFLOW_NAME_ONLY should move to the newer session-name controls.